Elamur Research⌕
☰
home / papers / 2026 / HarfoSokhan: A Comprehensive Parallel Dataset for Transitions between Persian Colloquial and Formal Variations

HarfoSokhan: A Comprehensive Parallel Dataset for Transitions between Persian Colloquial and Formal Variations

Paper · 2026 · Hamid Jahad Sarvestani, Vida Ramezanian, Saee Saadat et al.

Summary

Persian has a formal written register and a colloquial one ("Shekaste-nevisi"). The two differ in vocabulary, morphology and syntax. Models built for standard Persian lose accuracy on informal social-media text, and no public large-scale parallel corpus linked the two registers. The authors address this with HarfoSokhan, a colloquial-to-formal parallel dataset of about 6M sentence pairs. It combines two parts:

  • a small, high-quality core of roughly 12K pairs written and checked by native speakers;
  • a much larger synthetic portion made by back-translation, in which colloquial sentences go through English and return as formal Persian via Google Translate.

The authors checked that pipeline's output against human translators using METEOR. They trained models on the data, notably a GPT2 variant, and evaluated them with a ranking scheme in which native speakers order competing formalizations. GPT2-HarfoSokhan outperformed GPT-3.5-turbo and a leading rule-based converter. Its formalized text also gave better downstream ParsBERT classification results than ChatGPT output. The resource offers a practical basis for normalizing user-generated Persian and making existing NLP tools more robust.

Key points

  • The dataset has ~6M colloquial–formal pairs: a ~12K human-made core plus back-translated data with English as the pivot. It is released on Hugging Face (llm-lab/HarfoSokhan) under a Creative Commons NonCommercial license.
  • Pairwise METEOR between Google Translate and human annotators (averages roughly 0.45–0.49) was on par with agreement between the annotators themselves. This supported using machine translation to generate the large corpus.
  • In downstream tests with ParsBERT, text formalized by GPT2-HarfoSokhan beat text formalized by ChatGPT (GPT-4o):
    • news classification (7 classes): 74.60% vs 71.43% accuracy;
    • sentiment analysis (3 classes): 85.09% vs 83.33% accuracy.
  • Quality drops with sentence length. GPT2-HarfoSokhan's top@1 score is 50.0% for short sentences but only 28.5% once sentences exceed 150 characters.
  • The error analysis reports three recurring weaknesses: colloquial verb stems or suffixes left unconverted, irregular pronouns and slang only partly formalized, and proper nouns wrongly altered.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Hamad bin Khalifa UniversityDoha · Qatar141 works · 14,175 citations🎓University of BonnBonn · Germany10 works · 5,751 citations🎓Sharif University of TechnologyTehran · Iran5 works · 524 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

Cite

Cite as Hamid Jahad Sarvestani, Vida Ramezanian, Saee Saadat, Neda Taghizadeh Serajeh, Maryam Sadat Razavi Taheri, Shohreh Kasaei, MohammadAmin Fazli and Ehsaneddin Asgari. 2026. HarfoSokhan: A Comprehensive Parallel Dataset for Transitions between Persian Colloquial and Formal Variations. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eacl-long.346

@inproceedings{sarvestani2026harfosokhan,
  title = {{HarfoSokhan: A Comprehensive Parallel Dataset for Transitions between Persian Colloquial and Formal Variations}},
  author = {Hamid Jahad Sarvestani, Vida Ramezanian, Saee Saadat, Neda Taghizadeh Serajeh, Maryam Sadat Razavi Taheri, Shohreh Kasaei, MohammadAmin Fazli and Ehsaneddin Asgari},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-long.346},
  url = {https://doi.org/10.18653/v1/2026.eacl-long.346}
}

References

22 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Julien Abadji, Pedro Ortiz Suarez, Laurent Romary, and Benoît Sagot. 2022. Towards a cleaner document-oriented multilingual crawled corpus. arXiv preprint arXiv:2201.06642.
  2. Satanjeev Banerjee and Alon Lavie. 2005. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65–72, Ann Arbor, Michigan. Ass…
  3. Mehrdad Farahani, Mohammad Gharachorloo, Marzieh Farahani, and Mohammad Manthouri. 2021. Pars-BERT: Transformer-based model for Persian language understanding. Neural Processing Letters, 54:365–383. find
  4. FarsiYar Text Mining Group. 2020. FarsiYar text mining tools. https://github.com/Text-Mining? language=go. ↗
  5. Hooshvare Lab. 2021a. Persian news dataset. https://hooshvare.github.io/docs/datasets/tc. ↗
  6. Hooshvare Lab. 2021b. Persian sentiment analysis dataset. https://hooshvare.github.io/docs/ datasets/sa. ↗
  7. Hooshvare Team. 2021. ParsGPT-2, a Persian version of GPT-2. https://github.com/hooshvare/ parsgpt. ↗
  8. Mohsen Mahdavi Mazdeh. 2013. Colloquial Persian lessons. https://persiandee.com/view/ colloquial/currentLesson/0. ↗
  9. MUT Deep Group. 2022. Degarbayan-SC: A colloquial paraphrase Farsi subtitles dataset. https://github.com/mut-deep/Degarbayan-SC. ↗
  10. Amin Naemi, Marjan Mansourvar, Mostafa Naemi, Bahman Damirchilu, Ali Ebrahimi, and Uffe Kock Wiil. 2021. Informal-to-formal word conversion for Persian language using natural language processing techniques. In Proceedings of the 2nd International Conference on Computing, Networks and Internet of Thi… find
  11. OpenAI. 2022. ChatGPT. blog/chatgpt. https://openai.com/ ↗
  12. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA. Association for Computational…
  13. Ahmad Pouramini. 2021. parsT5-base. https://huggingface.co/Ahmad/parsT5-base. ↗
  14. Shrimai Prabhumoye, Yulia Tsvetkov, Ruslan Salakhutdinov, and Alan W. Black. 2018. Style transfer through back-translation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 866–876, Melbourne, Australia. Association for Computa…
  15. Internet Society Foundation. 2023. What are Alec Radford, Jeffrey Wu, Rewon Child, David Luan, the most used languages on the internet? Dario Amodei, and Ilya Sutskever. 2019. Language https://www.isocfoundation.org/2023/05/ models are unsupervised multitask learners. OpenAI what-are-the-most-used-l… ↗
  16. Roya Kabiri, Simin Karimi, and Mihai Surdeanu. 2022. Informal Persian universal dependency treebank. In Sudha Rao and Joel Tetreault. 2018. Dear sir or madam, Proceedings of the Thirteenth Language Resources may I introduce the GYAFC dataset: Corpus, bench-and Evaluation Conference (LREC 2022), page… find
  17. Tianxiao Shen, Tao Lei, Regina Barzilay, and Tommi Bruno Martins and Mário J. Silva. 2004. Spelling cor-Jaakkola. 2017. Style transfer from non-parallel text rection for search engine queries. In Proceedings by cross-alignment. In Proceedings of the 31st In-of the 4th International Conference on Adv… find
  18. Vahide Tajalli, Fateme Kalantari, and Mehrnoush Shamsfard. 2023. Developing an informal-formal Persian corpus. arXiv preprint arXiv:2308.05336. — [arXiv:2308.05336](https://arxiv.org/abs/2308.05336) find
  19. Jörg Tiedemann. 2012. Parallel data, tools and interfaces in OPUS. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12), Istanbul, Turkey. European Language Resources Association (ELRA). find
  20. University of Texas at Austin, Department of Middle Eastern Studies. 2024. Persian language program. https://liberalarts.utexas.edu/ languages/persian.html. Accessed: 2024-09-02. ↗
  21. Yunli Wang, Yu Wu, Lili Mou, Zhoujun Li, and Wen-Han Chao. 2020. Formality style transfer with shared latent space. In Proceedings of the 28th International Conference on Computational Linguistics, pages 2236–2249, Barcelona, Spain (Online). International Committee on Computational Linguistics. two-… find
  22. B Hazm and NLTK were used for text preprocessing. NLTK, METEOR, and Spacy were used for evaluation, with default parameter settings unless otherwise specified. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sarvestani-2026-harfosokhan-comprehensive-parallel-dataset-transitions/